acceptodds
Under review as a conference paper at ICLR 2027

PosterText: Unifying Text Patch Generation and Editing for E-Commerce Poster Design

Abstract

Automated e-commerce poster design requires not only the generation of high-quality posters from scratch but also the ability to flexibly and precisely edit existing designs. However, most existing methods either focus on end-to-end poster generation or follow multi-stage design pipelines, with limited capability for flexible and fine-grained editing of existing posters. To enable unified generation and editing of e-commerce posters, we introduce Text Patch Generation and Editing, a unified task formulation that treats text patches as atomic units and covers four fundamental operations: poster generation, patch addition, patch deletion, and patch modification, with optional reference-guided style condition control. Based on this formulation, we propose PosterText, a unified model trained with a four-stage curriculum, including text rendering pretraining, instruction-following training, reinforcement learning for preference alignment, and spatial guidance self-distillation for execution refinement. To facilitate systematic training and evaluation, we further construct a large-scale dataset with fine-grained patch-level annotations, together with a comprehensive benchmark covering both generation and editing scenarios. Extensive experiments demonstrate that PosterText achieves competitive performance against existing generation and editing approaches, validating the effectiveness of the proposed framework.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Poster Generation and Editing, Text Rendering, Diffusion Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_1diCOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PosterText, a unified framework for e-commerce poster generation and editing. It treats text patches as basic units and supports poster generation, patch addition, deletion, and modification within a single model. The model is trained with a four-stage curriculum including text rendering pretraining, instruction-following training, preference alignment, and Spatial Guidance Self-Distillation (SGSD). The paper also introduces PosterText-320K and PosterText-Bench for training and evaluation.

Critical strengths

1、Practical and useful task formulation. Unifying poster generation and localized text editing in a single framework is meaningful for real-world e-commerce poster design.

2、Solid implementation and evaluation. The paper provides a complete training pipeline, dataset, benchmark, and extensive experiments, with strong results in text fidelity, region preservation, and instruction following.

Critical weaknesses

1、The methodological novelty is relatively limited. The main technical pipeline largely combines existing components, including staged supervised training, DiffusionNFT-based preference alignment, and teacher-student self-distillation. The main contribution appears to lie more in task formulation and system integration than in a fundamentally new learning method.

2、The experiments do not directly validate the benefit of the unified generation-and-editing framework. Table 2 mainly studies the contribution of different training stages, but does not compare the unified model against separately trained generation-only and editing-only models. Therefore, the current ablation validates the training recipe rather than the necessity of unifying generation and editing.

Questions for authors

1、Can the authors compare PosterText with separately trained generation-only and editing-only models under comparable model capacity and training budgets?

2、Can the authors include or further discuss more task-specific visual text generation/editing baselines?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 1diC (part 1/2)

We appreciate the reviewer's valuable and constructive comments. Our detailed responses to each question and concern are provided below.

W1: Clarification of Methodological Contributions

We agree that individual technical components (staged SFT, DiffusionNFT, and teacher-student distillation) build upon existing works. Rather than claiming novelty on these standalone modules, our core contributions center on three aspects:

  1. Task Formulation. We formulate stylized text patches (comprising text content, typography style, decorative backing, and spatial region) as the shared fundamental unit for both generation and editing. This unifies poster generation alongside patch addition, deletion, and modification under a single interface with patch-level reference style control. In contrast, existing poster-specific methods operate globally at the poster or canvas level, making local edits such as replacing a single slogan or deleting outdated text inconvenient, as they often force full-poster regeneration or require user-provided masks and bounding boxes.
  2. Training Curriculum for E-Commerce Posters. We design a four-stage curriculum targeting the typical failure modes of e-commerce posters: Stage I establishes glyph rendering, Stage III introduces a background preservation reward to constrain editing boundaries, and SGSD supervises the student along its own rollout trajectories using a mask-conditioned teacher, enabling mask-free localization at inference time. Ablation studies verify the effectiveness of each stage.
  3. Empirical Gains from Unified Modeling. Controlled experiments comparing joint and task-specific training (Appendix E) reveal clear bidirectional transfer between generation and editing in our text-centric setting, showing that unified modeling yields measurable performance gains alongside practical convenience.

W2 & Q1: Comparison of Joint and Task-Specific Training

To verify the effectiveness of joint training, we conducted a controlled comparison between joint and task-specific training. All variants start from the same Stage I checkpoint and share identical Stage II configurations, including LoRA rank, batch size, learning rate, and schedule, as well as the same underlying data pool, ensuring matched model capacity. The variants differ only in their training data composition and task partitioning:

  • Unified: Jointly trains a single LoRA across all four tasks with balanced task sampling for 10K steps
  • Gen-only / Edit-only: Trains solely on generation data or solely on editing data (addition, deletion, and modification) for 10K steps each, serving as equal-step controls corresponding to the generation-only and editing-only models mentioned by the reviewer.
  • 4× Single-task: Trains a separate LoRA for each task for 2.5K steps each, matching the total training step budget of the Unified model.

Table R3-1: Comparison of joint and task-specific training.

ModelStepsGen S.A.↑Gen NED↑Gen B.P.↑Add S.A.↑Add NED↑Mod S.A.↑Mod NED↑Del M-L.↓Del I.F.↑
Gen-only10K0.7530.8124.61––––––
Edit-only10K–––0.6970.8490.7120.8660.0564.09
4× Single-task2.5K each0.7540.8194.580.6840.8310.7020.8570.0624.08
Unified10K0.7580.8204.740.7060.8610.7180.8710.0584.11

Within our text-centric e-commerce poster setting, Unified achieves competitive or superior results compared to both the equal-step Gen-only and Edit-only controls as well as the four single-task models under the same total compute budget, while requiring only a single adapter. This demonstrates clear bidirectional transfer between generation and editing without noticeable negative transfer. Specifically, editing supervision enforces strict preservation of the product and non-target regions, transferring to the generation task to yield a higher background preservation score. Conversely, generation requires rendering all text across the entire poster from scratch with broader glyph coverage, thereby improving both Sen.Acc and NED on text-rendering edits such as addition and modification. For text deletion, which does not involve text rendering, differences across variants remain minimal. More details are provided in Appendix E.

AuthorsOpenReview ↗

Response to Reviewer 1diC (part 2/2)

Q2: Comparison with Additional Baselines

We have added three poster-specific methods (PosterCraft, PosterMaker, and PosterOmni) and two general-purpose editing methods (LongCat-Image-Edit and Boogu-Image-Edit), while supplementing the generation evaluation with text fidelity metrics (Sen.Acc and NED). Since none of the three poster-specific methods directly accept our input interface, we apply minimal adaptations to each baseline:

  • PosterCraft: Generates posters starting from the product image via an SDEdit-style pipeline.
  • PosterMaker: Incorporates a Qwen3.5-9B planner to predict scene descriptions and text bounding boxes from the input instruction and target image dimensions.
  • PosterOmni: Routes generation, addition, and modification through its ID-driven branch, and handles text deletion via its mask-based filling branch. Since PosterCraft and PosterMaker do not support editing existing posters, they are evaluated solely on the generation task.

Table R3-3: Results on the poster generation task.

MethodSen.Acc↑NED↑I.F.↑V.Q.↑B.P.↑
LongCat-Image-Edit0.7490.8113.794.284.12
Boogu-Image-Edit0.7610.8173.954.163.38
PosterCraft0.5870.6753.233.862.52
PosterMaker0.7310.7983.623.754.76
PosterOmni0.7230.7923.844.213.04
PosterText0.8020.8684.214.464.88

Table R3-4: Results on patch editing tasks.

MethodAdd Sen.Acc↑Add M-L.↓Del M-L.↓Mod Sen.Acc↑Mod M-L.↓
LongCat-Image-Edit0.3540.1280.1040.6580.198
Boogu-Image-Edit0.3690.3170.0890.6910.194
PosterOmni0.3890.4310.1850.5940.217
PosterText0.7630.0390.0400.7710.028

For the generation task, PosterText achieves the best performance across all metrics among the newly added baselines. While PosterMaker preserves the product well, it depends on an external planner to translate instructions into text bounding boxes, so any planning error directly degrades text placement or styling. Meanwhile, PosterCraft and PosterOmni lack explicit subject-preservation mechanisms and are therefore more prone to altering the product. For editing tasks, PosterText achieves substantially stronger text modification fidelity and background preservation. In contrast, PosterOmni relies on global image-level or canvas-level operations and tends to recompose the entire scene during text addition and modification rather than locally updating a single specified line, resulting in considerably larger unintended alterations in non-target regions. Complete experimental results are provided in Table 1.

We hope these additions address your concerns, and we kindly ask the reviewer to take these results into consideration in your final evaluation. We would be glad to discuss further if any additional questions arise.

AuthorsOpenReview ↗

Gentle Reminder: Author Response Ready for Review

We sincerely appreciate your valuable time and constructive feedback! We hope our previous response and revisions have adequately addressed your concerns. As the discussion period is about to close very soon, we kindly invite you to review our response at your earliest convenience. We warmly welcome any remaining questions and would be delighted to provide further clarification before the deadline!

Thanks for the detailed response. Authors addressed my concerns and I raise my score to 3.

Reviewer_6Zu8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PosterText uses styled text patches to unify poster generation, addition, deletion, and modification. A Qwen-Image-Edit backbone undergoes rendering pretraining, instruction tuning, reward alignment, and spatial guidance self-distillation (SGSD), where a mask-conditioned EMA teacher supervises a mask-free student. The paper introduces PosterText-320K and a 750-record benchmark.

Critical strengths
  • Editing text together with its decoration matches a practical need that character replacement alone misses. Figures 1, 4, and 6 show localized, reference-guided edits while preserving surrounding content.
  • Addition sentence accuracy is 0.763 versus 0.543 for the best listed closed-source result, with NED of 0.899 versus 0.679 (Table 1). Low masked LPIPS indicates that these gains do not require redrawing the entire poster.
  • Stage ablations improve sentence accuracy from 0.679 to 0.712 and then 0.767 (Table 2). The SGSD student also approaches the mask-guided teacher, 0.767 versus 0.781, and exceeds the pre-distillation model's 0.735 (Appendix B.4).
Critical weaknesses
  • Table 1 omits poster-specialized methods such as PosterCraft, PosterMaker, POSTA, and PosterOmni, despite citing all four. General-purpose models are useful references, but do not establish progress over prior poster systems. Restricting baselines to methods supporting both generation and editing is unnecessarily limiting: compare specialists on compatible tasks, with product, layout, and mask conditioning explicitly accounted for.
  • Section 2.1 overstates the lack of controllable editing in prior poster methods. PosterOmni already combines local editing with global creation, while POSTA supports layout, typography, and style customization. Neither is identical to text-patch editing, but the paper should distinguish its patch representation and SGSD from these capabilities rather than frame generation-editing unification itself as the unresolved gap.
  • SGSD adds 500 optimization steps without an equal-step continuation baseline, mask-free teacher control, or teacher-forced comparison. The ablation supports the full recipe but does not isolate the value of privileged masks and student-trajectory distillation.
  • The 750-record benchmark lacks counts by operation, language, patch complexity, and reference usage. Deduplication does not establish transfer beyond the synthetic construction process; unseen products/templates or independently authored instructions would provide a stronger test.
  • Human-judge correlations of 0.80/0.74/0.82 lack sample counts and annotation details. Table 5 also labels its averages editing-only but repeats all-task values: Table 1's editing I.F. scores average to 4.58, not 4.49. This appears to be a reporting error, not a failure of the method.
Questions for authors
  1. Does SGSD outperform equal-step SFT/RL continuation and distillation from a mask-free teacher?
  2. Why are the cited poster-specialized methods absent from Table 1? Please compare PosterCraft, PosterMaker, POSTA, and PosterOmni on compatible task subsets, account for their different conditioning requirements, and include text-fidelity metrics for full-poster generation.
  3. How many outputs and annotators support the judge audit, and with what uncertainty? Please also reconcile Table 5's editing-only label with its reported averages.
4 comments
AuthorsOpenReview ↗

Response to Reviewer 6Zu8 (part 1/3)

We sincerely thank the reviewer for the careful and constructive comments to improve the quality of our paper. Below, we provide our detailed responses to each of the questions and concerns raised.

W1 & Q2: Comparison with Poster-Specific Methods

We have added three poster-specific baselines (PosterCraft, PosterMaker, and PosterOmni) to Table 1, and supplemented the generation evaluation with text fidelity metrics (Sen.Acc and NED). Since POSTA has not released its model weights and thus cannot be reproduced, we limit our discussion of it to a capability-level comparison.

Because these methods differ in their required conditional inputs, we apply minimal adaptations to align them with our task interface:

  • PosterCraft: As it accepts only text prompts, we adopt an SDEdit-style pipeline initialized from the product image, encoding the target text, font, color, and layout requirements into the prompt.
  • PosterMaker: We equip it with a Qwen3.5-9B planner to predict scene descriptions and text bounding boxes from the input instruction and target image dimensions.
  • PosterOmni: It handles text deletion via its filling branch conditioned on the target mask, while routing the remaining tasks through its ID-driven branch using the source image, prompt, and reference patches as inputs.

Table R2-1: Results on the poster generation task.

MethodSen.Acc↑NED↑I.F.↑V.Q.↑B.P.↑
PosterCraft0.5870.6753.233.862.52
PosterMaker0.7310.7983.623.754.76
PosterOmni0.7230.7923.844.213.04
PosterText0.8020.8684.214.464.88

Table R2-2: Results on poster editing tasks.

MethodAdd Sen.Acc↑Add M-L.↓Add I.F.↑Del M-L.↓Del I.F.↑Mod Sen.Acc↑Mod M-L.↓Mod I.F.↑
PosterOmni0.3890.4313.670.1853.860.5940.2173.34
PosterText0.7630.0394.510.0404.700.7710.0284.53

On the generation task, PosterText achieves the best performance across metrics among the poster-specific methods. While PosterMaker preserves the foreground product well by fixing the subject region during generation, it strictly requires pre-defined text bounding boxes that must be predicted from instructions via an external planner. Any planning errors directly translate into placement or styling mistakes, meaning its results reflect the combined quality of both the generator and the planner. In contrast, PosterCraft lacks explicit product-preservation conditioning, and PosterOmni's ID-driven branch tends to recompose the entire layout, making both models more prone to altering the product.

On editing tasks, the performance gap becomes even more pronounced. Because PosterOmni operates globally across the entire image or canvas, it tends to recompose the whole scene rather than locally modifying a single specified line of text, leading to substantial unintended alterations in non-target regions.

W2: Differences from Existing Poster Methods

We thank the reviewer for this feedback. We agree that framing unified generation and editing in overly broad terms fails to properly distinguish our scope from prior works, so we have revised Section 2.1. Specifically, while POSTA supports multi-stage layout, typography, and style customization, and PosterOmni enables canvas-level operations (creation, filling, outpainting, and resizing) alongside reference-driven generation, both target creative poster design with a focus on global layout and overall visual style. Because their controls operate at the full-poster or canvas level, they are prone to unintended layout or background drift when applied to fine-grained local edits. In contrast, our e-commerce setting is text-centric: the product must remain intact while everyday workflows require locally replacing a slogan, restyling a price tag from a reference, or deleting outdated text without altering surrounding regions. Accordingly, we formulate stylized text patches as the fundamental unit for both generation and editing, enabling patch-level style control and handling all four operations via a single natural language instruction at inference time.

AuthorsOpenReview ↗

Response to Reviewer 6Zu8 (part 2/3)

W3 & Q1: Ablation Study of SGSD

To validate the effectiveness of SGSD, we conducted an additional set of controlled experiments. All variants are initialized from the same Stage III checkpoint and trained for an additional 500 steps, varying only the training objective or supervision strategy:

  • (a) Continued SFT: Continues flow matching training without masks, using the GT velocity as the target.
  • (b) Continued NFT: Continues DiffusionNFT training with the same reward functions and hyperparameters used in Stage III.
  • (c) Mask-free Teacher Distillation: Shares the exact same architecture as SGSD, except that the teacher receives mask-free conditions.
  • (d) Teacher-forced Distillation: Uses a mask-conditioned teacher, but computes supervision on states obtained by adding noise to GT images.
  • (e) SGSD: Uses a mask-conditioned teacher and supervises directly along the student's self-generated rollout trajectory.

Table R2-3: Ablation results of SGSD.

VariantS.A.↑NED↑M-L.↓I.F.↑V.Q.↑
Stage III0.7510.8760.0374.224.51
(a) + Continued SFT0.7430.8680.0394.184.46
(b) + Continued NFT0.6830.7960.0943.893.88
(c) + Mask-free Teacher Distillation0.7560.8810.0384.244.54
(d) + Teacher-forced Distillation0.7700.8830.0364.414.55
(e) + SGSD0.7790.8900.0364.494.62

Under the same budget, continued SFT yields no improvement because the reconstruction objective has already converged in Stage II. Continued NFT degrades performance across all metrics, indicating RL over-optimization and reward hacking under sparse terminal rewards. Removing mask conditioning from the teacher yields only marginal gains. Building upon mask conditioning, supervision along student trajectories accounts for the remaining improvements: teacher-forced distillation only supervises states along noisy GT paths, meaning the student never encounters its own off-track sampling states and thus fails to learn error correction, resulting in limited gains. In contrast, SGSD directly rectifies errors along the student's own trajectory, achieving substantial gains in I.F. and V.Q. More details are provided in Appendix F.

W4: Benchmark Statistics and Out-of-Distribution Generalization

Benchmark Statistics. We have added detailed benchmark statistics in Appendix D.3, categorized across four dimensions: operation type, language, reference patch usage, and the number of target patches.

Table R2-4: Statistics of PosterText-Bench.

MetricAddModDelGen
Number of samples200200150200
Chinese / English150 / 50150 / 50120 / 30150 / 50
With reference patch12010000
Avg. target patches2.82.11.45.6

Out-of-Distribution Evaluation. To better evaluate model generalization, we constructed a new out-of-distribution test set of 200 samples (50 per task). The images are sourced from real-world e-commerce posters and merchant product photos, and are strictly deduplicated against PosterText-320K. All instructions are manually written and annotated by human annotators viewing the images, completely independent of our automated data construction pipeline.

Table R2-5: Results on the out-of-distribution data.

MethodS.A.↑NED↑M-L.↓I.F.↑V.Q.↑
Seedream 5.00.6130.7220.1134.314.59
Nano Banana Pro0.6680.7730.0924.484.64
FireRed-Image-Edit0.4630.6210.3423.693.82
Qwen-Image-Edit0.4720.6500.4273.143.31
PosterText0.7170.8110.0464.344.51

On out-of-distribution data, PosterText achieves competitive advantages in S.A., NED, and M-L., demonstrating that its text rendering accuracy and non-target region preservation generalize effectively to unseen posters and human-written instructions. While it remains slightly behind the closed-source baselines on I.F. and V.Q., it outperforms all open-source baselines by a substantial margin across both metrics. We attribute the remaining gap on I.F. primarily to the shift in instruction style: human-written instructions refer to target elements in much more diverse ways, whereas closed-source models possess stronger language backbones and are thus less sensitive to such stylistic variations. Incorporating more human-written or paraphrased instructions during training should help close this gap. Further details are provided in Appendix G.

AuthorsOpenReview ↗

Response to Reviewer 6Zu8 (part 3/3)

W5 & Q3: Human Verification Details and Table 5 Metric Protocol

Human Verification. We have added the details of human verification in Appendix D.2. Three annotators with e-commerce design experience independently scored 750 outputs produced by our model on PosterText-Bench using the exact same 1 to 5 rating rubric as the VLM evaluator, with the human score for each output taken as the average across the three annotators. We then computed the Spearman rank correlation coefficients between the human scores and the VLM evaluator scores, obtaining 0.80 for I.F., 0.74 for V.Q., and 0.82 for B.P.

Metric Protocol. We thank the reviewer for catching this carefully. This was a labeling error in the table caption rather than an issue with the reported numbers themselves. Table 5 follows the exact same averaging protocol as Table 2: S.A. and NED are averaged over the three text-rendering tasks (addition, modification, and generation), M-L. is averaged over the three editing tasks, and I.F. and V.Q. are averaged across all four tasks. We have updated the table caption and annotations accordingly to ensure consistency.

We hope these additions address the reviewer's questions, and we would be happy to discuss further if any additional concerns arise.

AuthorsOpenReview ↗

Gentle Reminder: Author Response Ready for Review

We sincerely appreciate your valuable time and constructive feedback! We hope our previous response and revisions have adequately addressed your concerns. As the discussion period is about to close very soon, we kindly invite you to review our response at your earliest convenience. We warmly welcome any remaining questions and would be delighted to provide further clarification before the deadline!

Reviewer_RxvAOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the disconnect between generation from scratch and local editing in e-commerce posters, and introduces the task of Text Patch Generation and Editing. It treats a text patch as the fundamental unit, enabling a single model to support poster generation, patch addition, patch deletion, and patch modification. Both patch addition and modification can be guided by reference patches for style control.

Critical strengths
  1. The research problem is well-motivated and practically valuable. E-commerce poster design workflows do require both initial generation and repeated local editing. Treating text as a structured visual unit with style, decoration, and spatial attributes is more consistent with real-world design objects.
  2. The training design is comprehensive, with each component reasonably aligned with the corresponding failure modes. Stage I addresses glyph accuracy, Stage II learns the task interface, and Stage III optimizes text quality, aesthetics, and background preservation. SGSD specifically targets localization and execution errors during inference without masks.
  3. The evaluation coverage is relatively comprehensive. PosterText-Bench covers all four task types. The appendix also reports the Spearman correlation between VLM-based judgments and human annotations.
Critical weaknesses
  1. A key ablation of unified training itself is missing. Table 2 compares different training stages, but does not compare joint training on all four tasks against four task-specific models (or single-task LoRA models trained with the same total data and compute budget). Therefore, the experiments demonstrate the effectiveness of the final unified model, but do not establish that unified learning avoids significant negative transfer, outperforms separate training, or that editing-oriented control genuinely benefits generation.
  2. The baseline selection provides limited evidence of competitiveness in full poster generation. The paper compares only methods that support both generation and editing, omitting strong specialized poster generation systems discussed in Related Work (e.g., PosterMaker, PosterCraft, and PosterOmni) as well as dedicated text editing methods. This choice helps validate the claim that a single interface can cover all four tasks, but is insufficient to support broader claims about generation competitiveness.
Questions for authors

I understand that a unified task formulation does seem more elegant, but I would still like to see how joint training on all tasks compares with task-specific training under the same backbone and LoRA capacity.

4 comments
AuthorsOpenReview ↗

Response to Reviewer RxvA

We thank the reviewer for the constructive and thoughtful feedback. Below, we address each question and concern point by point.

W1 & Q1: Comparison between Joint and Separate Training

To verify the effectiveness of joint training, we conducted an additional set of controlled experiments comparing joint and task-specific training. All variants start from the same Stage I checkpoint and share identical Stage II training configurations, including LoRA rank, batch size, learning rate, schedule, and the underlying data pool. The variants differ only in their training data composition and task partitioning:

  • Unified: Jointly trains a single LoRA across all four tasks with balanced task sampling for 10K steps.
  • 4× Single-task: Trains a separate LoRA for each task for 2.5K steps each, matching the total training step budget of the Unified model.
  • Gen-only / Edit-only: Trains solely on generation data or solely on editing data (addition, deletion, and modification) for 10K steps each. These two variants serve as equal-step controls to examine cross-task transfer between generation and editing.

Table R1-1: Comparison of joint and task-specific training.

ModelStepsGen S.A.↑Gen NED↑Gen B.P.↑Add S.A.↑Add NED↑Mod S.A.↑Mod NED↑Del M-L.↓Del I.F.↑
Gen-only10K0.7530.8124.61––––––
Edit-only10K–––0.6970.8490.7120.8660.0564.09
4× Single-task2.5K each0.7540.8194.580.6840.8310.7020.8570.0624.08
Unified10K0.7580.8204.740.7060.8610.7180.8710.0584.11

Within our text-centric e-commerce poster setting, Unified outperforms all four single-task models across all average metrics under the same total compute budget, while requiring only a single adapter. Furthermore, the equal-step controls demonstrate clear bidirectional transfer between generation and editing without noticeable negative transfer: editing supervision enforces strict preservation of the product and non-target regions, yielding a higher background and product fidelity score on the generation task; conversely, generation requires rendering all text across the entire poster from scratch with broader glyph coverage, thereby improving both Sen.Acc and NED on text-rendering edits such as addition and modification. For text removal, which does not involve text rendering, the differences across variants remain minimal. More results and analysis are provided in Appendix E.

W2: Comparison with Additional Baselines

We further include three poster-specific methods (PosterCraft, PosterMaker, and PosterOmni) and extend the generation evaluation with text fidelity metrics (Sen.Acc and NED). Detailed results are presented in Table 1. Since PosterCraft and PosterMaker cannot natively accept our task inputs (product image, instruction), we introduce minimal adaptations for fair evaluation:

  • PosterCraft: As it accepts only text prompts, we adopt an SDEdit-style pipeline initialized from the product image, encoding the target text, font, color, and layout requirements into the prompt.
  • PosterMaker: We equip it with a Qwen3.5-9B planner to predict scene descriptions and text bounding boxes from the input instruction and target image dimensions.

Table R1-2: Results of Additional Methods on Poster Generation

MethodSen.Acc↑NED↑I.F.↑V.Q.↑B.P.↑
PosterCraft0.5870.6753.233.862.52
PosterMaker0.7310.7983.623.754.76
PosterOmni0.7230.7923.844.213.04
PosterText0.8020.8684.214.464.88

On the generation task, PosterText achieves the best performance across all five metrics among poster-specific baselines. While PosterMaker preserves the product well, its text rendering relies on an external planner to convert instructions into bounding boxes first, meaning planning errors directly cause placement or styling mistakes. Meanwhile, PosterCraft and PosterOmni lack explicit subject-preservation mechanisms and are therefore more prone to altering the product. Regarding dedicated text-editing methods, similar to PosterMaker, they require user-specified masks and operate on raw text strings rather than stylized patches, resulting in a distinct task interface from our setting. More quantitative results are presented in Table 1, and adaptation details are provided in Appendix A.3.

We hope the above results address the reviewer's concerns, and we would be happy to provide further details if needed.

AuthorsOpenReview ↗

Gentle Reminder: Author Response Ready for Review

We sincerely appreciate your valuable time and constructive feedback! We hope our previous response and revisions have adequately addressed your concerns. As the discussion period is about to close very soon, we kindly invite you to review our response at your earliest convenience. We warmly welcome any remaining questions and would be delighted to provide further clarification before the deadline!

Thank you for the authors' response. My concerns have been addressed, particularly regarding the improvements of joint training over single-task training. I will therefore keep my score unchanged.

Thank you very much for your follow-up and for your positive assessment of our work. We sincerely appreciate your constructive feedback throughout the review process, which has helped us improve the clarity and quality of the manuscript. We are happy to discuss any further questions or suggestions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Change Summary

We appreciate all valuable feedback and comments from reviewers. The revised manuscript includes the following key updates, with all modifications marked in blue for clarity:

  1. Additional baselines (Reviewer RxvA, Reviewer 6Zu8, Reviewer 1diC). Table 1 now includes three poster-specialized methods (PosterCraft, PosterMaker, and PosterOmni) and two general-purpose editing models (LongCat-Image-Edit and Boogu-Image-Edit). For poster generation, we also report the text fidelity metrics S.A. and NED. Section 4 discusses the specialized methods, and Appendix A.3 describes how each baseline was adapted to our setting.
  2. Positioning against prior poster methods (Reviewer 6Zu8, Reviewer 1diC). We rewrote the relevant part of the Introduction to clarify how PosterText differs from POSTA and PosterOmni. Those methods mainly target full-poster generation or canvas-level operations. PosterText instead treats each styled text patch as the basic unit and performs fine-grained, mask-free editing from natural-language instructions. The Conclusion now limits our scope to e-commerce posters.
  3. Joint vs. separate training (Reviewer RxvA, Reviewer 1diC). New Appendix E compares Unified, Gen-only, Edit-only, and 4× Single-task training under matched compute budgets. The results show bidirectional transfer between generation and editing.
  4. SGSD ablation (Reviewer 6Zu8). New Appendix F compares continued SFT, continued NFT, mask-free teacher distillation, teacher-forced distillation, and SGSD, all with the same number of extra training steps. The gains come from the distillation mechanism, not from additional optimization. Appendix B.4 adds an analysis of mask-guided spatial transfer.
  5. Human verification, benchmark statistics, and OOD evaluation (Reviewer 6Zu8).
    • Appendix D.2 details how we validated the VLM judge against human annotations.
    • Appendix D.3 adds statistics for PosterText-Bench.
    • New Appendix G evaluates generalization on 200 out-of-distribution samples with human-written instructions.
  6. Metric protocol correction (Reviewer 6Zu8). We fixed a labeling error in the Table 5 caption. S.A. and NED are now consistently averaged over the addition, modification, and generation tasks, and we updated the numbers and the corresponding text to match.

These updates enhance the clarity of our scope and positioning as well as the rigor and comprehensiveness of our experimental evaluation. We sincerely thank the reviewers again for their constructive comments.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.